For efficiency, the cardinality of the bloom filter
(i.e. the number of partition keys which will be written into the sstable)
has to be known before elements are inserted into the filter.
In some cases (e.g. memtables flush) this number is known exactly.
But in others (e.g. repair) it can only be estimated,
and the estimation might be very wrong, leading to an oversized filter.
Because of that, some time ago we added a piece of logic
(ran after the sstable is written, but before it's sealed)
which looks at the actual number of written partitions,
compares it to the initial estimate (on which the size of the bloom
filter was based on), and if the difference is unacceptably large,
it rewrites the bloom filter from partition keys contained in Index.db.
But the idea to rebuild the bloom filters from index files
isn't going to work with BTI indexes, because they don't store
whole partition keys. If we want sstables which don't have Index.db
files, we need some other way to deal with oversized filters.
Partition keys can be recovered from Data.db,
but that would often be way too expensive.
This patch adds another way. We introduce a new component file,
TemporaryHashes. This component, if written at all,
contains the 16-byte murmur hash for every partition key, in order,
and can be used in place of Index to reconstruct the bloom filter.
(Our bloom filters are actually built from the set of murmur hashes of
partition keys. The first step of inserting a partition key into a
filter is hashing the key. Remembering the hashes is sufficient
to build the filter later, without looking at partition keys again.)
As of this patch, if the Index component is not being written,
we don't allocate and populate a bloom filter during the Data.db write.
Instead, we write the murmur hashes to TemporaryHashes, and only
later, after the Data write finishes, we allocate the optimal-size,
bloom filter, we read the hashes back from TemporaryHashes,
and we populate the filter with them.
That is suboptimal.
Writing the hashes to disk (or worse, to S3) and reading
them back is more expensive than building the bloom filter
during the main Data pass.
So ideally it should be avoided in cases where we know
in advance that the partition key count estimate is good enough.
(Which should be the case in flushes and compactions).
But we defer that to a future patch.
(Such a change would involve passing some flag to the sstable writer
if the cardinality estimate is trustworthy, and not creating
TemporaryHashes if the estimate is trustworthy).
b1984d6798